Hard technology remains the focal point of technological competition, with semiconductor chips captivating the attention of people both nationally and globally.

Little known to many, alongside SoC 5G chips and AI-specific chips, there exists another promising type of semiconductor chip—the spectral chip.

In our physical world, when light interacts with matter, it produces a unique spectrum, akin to human fingerprints. Each substance possesses distinct spectral characteristics, and even a single crop can exhibit varying spectra at different growth stages. Through spectrometry, the spectral information of an object can be captured for detailed research and analysis.

A spectral chip is essentially a sensor comprising millions of microspectrometers. Leveraging the photoelectric effect of semiconductor materials, it can capture spectral information, previously only attainable through optical precision instruments, on a chip the size of a postage stamp.

From the laboratory to the field, the path to civilian applications for spectral chips shines brightly with AI. This article delves into how AI has transformed these chips, paving the way for spectral technology to enter the consumer market.

The transition from research-level technology to the civilian consumer market has never been an easy feat. Spectral technology, once a "highland flower" confined to laboratories, found little application in the civilian market.

Two primary factors hindered the civilian potential of spectral technology:

First, the cost. Traditional spectral information collection necessitates the use of bulky and expensive spectrometers, costing tens of thousands or even hundreds of thousands of yuan each. This price tag renders them unaffordable for ordinary households and low-profit industries like agriculture.

Second, ineffectiveness. Spectral analysis was primarily confined to laboratories, benefiting from a clean research environment and low signal-to-noise ratios. However, in open civilian environments, it faces numerous interferences, leading to poor signal-to-noise ratios and consequently affecting the efficiency and accuracy of spectral information analysis. Therefore, the physical solutions employed in laboratories were ineffective in civilian scenarios.

In the era of AI, the complementary nature of AI technology and spectral chips has finally made spectral technology accessible to the civilian market.

On one hand, the hardware handles modulation, while AI takes charge of demodulation, addressing the "ineffectiveness" issue in traditional spectral analysis. The underlying layer of the spectral chip is CMOS. With spectral modulation technology, one can discern the spectral composition of incident light, which AI algorithms then process and analyze, automatically extracting features from spectral data, filtering out noise, and performing baseline corrections. Thus, the images and spectral information obtained by consumer-grade devices are essentially "calculated" by AI.

On the other hand, the fusion of spectral physical information with machine vision digital information feeds AI, solving the "unaffordability" problem in industrial intelligence.

The performance of AI algorithms heavily relies on the scale and quality of data. Many industries lack sufficient information, necessitating economies of scale to enhance model performance, which demands high computational power and incurs costs. Spectral information, characterized by non-destructive testing, high efficiency, and accuracy, simplifies models, reducing complexity and computational demands, thereby easing application deployment.

It can be argued that the synergy between AI technology and spectral technology paves the way for widespread spectral chip applications in the consumer market.

Spectral chips boast a wide array of application scenarios. Among them, the consumer electronics field, with its vast audience and high commercial value, stands out as one of the most promising areas. Currently, AI-enhanced spectral chips have explored numerous applications in the consumer electronics market:

1. High-end mobile phones: Emphasizing imaging capabilities, high-end flagship phones utilize spectral chips to emit various colors, enhancing color options and fidelity in lighting and display fields.

2. Smart homes: Spectral chips are increasingly incorporated into smart home products like refrigerators, robotic vacuum cleaners, smart door locks, and smart cameras. For instance, refrigerators can detect food freshness using spectral cameras and AI models, reminding users to replace ingredients. Projectors leverage spectral cameras and color correction algorithms to mitigate color distortion caused by wall colors. In smart security, spectral chips enable rapid object recognition and tracking, as the collected spectral information is purely physical, resistant to deception by fake wigs or masks, fostering smarter and more secure smart door locks. Robotic vacuum cleaners employ spectral chips and recognition algorithms to precisely determine stain conditions and floor materials, intelligently adjusting cleaning strategies.

3. Automotive electronics: With the rise of intelligent vehicles, smart cockpits and autonomous driving are key areas of competition. Spectral chips play a pivotal role in both.

In smart cockpits, spectral cameras monitor the driver's safety and health, enhancing in-car security. In autonomous driving, spectral information accurately reflects road obstacles like foam, stones, and roadblocks, enabling the intelligent driving system to adopt distinct judgment strategies.

4. Low-altitude economy: AI spectral cameras combined with drones capture more accurate color information during low-altitude photography and videography, resulting in superior shooting outcomes.

5. Other applications: AI spectral chips can also be integrated with wearable devices for skin health monitoring.

Spectral chips exhibit immense potential not only in the consumer electronics market but also in the B-end industry market as digital transformation deepens across various industries. Spectral information, a highly reliable data type, accurately reflects crucial physical world indicators, increasing its application value.

Some institutions estimate the spectral chip industry's scale at approximately ten to one hundred billion yuan. However, from design to tape-out, mass production, and eventual commercial use, the journey is lengthy and fraught with uncertainty. What challenges must an AI spectral chip endure?

The head of Jilin Qiushi Spectrum revealed that they pioneered the "OCF spectral modulation + algorithm demodulation" technological innovation in 2017 and successfully taped out their first spectral chip in 2019. They now collaborate with leading domestic mobile phone manufacturers and home appliance companies, with AI integration as a key strategic direction.

The primary challenge in refining an AI spectral chip lies in funding.

The semiconductor chip industry is resource-intensive, heavily reliant on capital and talent, with high innovation risks. Most private capital shies away, historically relying on government subsidies and project systems. After Qiushi Spectrum's establishment, it secured investments from industry funds like Star Capital, Jilin Science and Technology Investment, and Changxing Fund, alleviating funding concerns for startups.

The next hurdle is data acquisition.

Spectral data stems from the physical world, making collection challenging and costly. Existing open-source datasets are inadequate, necessitating most data to be freshly collected. Currently, during spectral chip and industry solution development, Qiushi Spectrum has amassed substantial spectral data for model training based on project requirements. Future plans include collecting data from southern regions, extreme environments, dusty weather, etc., anticipating a data scale 4-5 times the current amount.

As data scales increase, so does the computational power challenge.

Initially, the company relied on expensive overseas N-cards for computing power, plagued by maintenance issues, limited video memory, and graphics card crashes, significantly impeding R&D efficiency. However, with local digital infrastructure upgrades and the province's first AI computing center, accessing the Changchun Computing Center's computing power services enabled multi-task parallel development without queuing. Previously, model optimization on an N-card took 2-3 days; now, results are available the next morning after overnight uploads, vastly improving R&D efficiency.

Manufacturing processes, however, pose less of a challenge. According to the head, spectral chips primarily utilize mature processes like 28nm, 40nm, and 55nm, which domestic manufacturing capabilities can support for large-scale mass production. Hence, after successfully taping out spectral chips in 2019, they swiftly entered the commercial stage. Leading manufacturers' mobile phones now offer more realistic imaging effects thanks to spectral chips.

In conclusion, a commercially viable AI spectral chip is supported by innovative scientific and technological investment models, consolidated semiconductor and optical industry capabilities, and upgraded digital infrastructure. Breakthroughs in hard technology do not happen overnight but are the culmination of deep roots and lush foliage.

When AI illuminates spectral chips, the fusion of the digital and physical worlds becomes infinitely possible.